Sign in

Duarte

@duarteocarmo.com
987 followers 232 following 424 posts

Born in 🇵🇹, based in 🇩🇰. Machine Learning Engineer (or whatever it's called now) duarteocarmo.com duarteocarmo.com/photos github.com/duarteocarmo/

PostsRepliesMedia
Duarte @duarteocarmo.com · 01/10/2026
LLM as a judge is a great idea!
210
Duarte @duarteocarmo.com · 29/09/2026
I decided to give Cursor another try today. I like this ngl
000
Duarte @duarteocarmo.com · 24/09/2026
Had to use cursor today (sshing into hpc - and wanted to get an idea of files, directories, etc) One thing Cursor still nails is the review pane. Lazygit and hunk are good - but this is better.
020
Duarte @duarteocarmo.com · 24/09/2026
I hate you 😂
130
Duarte @duarteocarmo.com · 24/09/2026
Am I the only one that is excited about this bad boy for local use? 9B - distilled from the #1 open weight model at the moment :) huggingface.co/ggml-org/Mi...
huggingface.co
ggml-org/MiMo-V2.6-Distill-Qwen-9B-GGUF · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
110
Duarte @duarteocarmo.com · 22/09/2026
That's what i thought
000
Duarte @duarteocarmo.com · 18/09/2026
Increasingly telling models that do "long horizon tasks"
000
Duarte @duarteocarmo.com · 17/09/2026
You use @herdrdev + pi - but sometimes the agent asks you a question and herdr does not show it as blocked github.com/duarteocarm...
github.com
dotfiles/.pi/agent/extensions/herdr-ui-prompt-bridge.ts at master · duarteocarmo/dotfiles
Contribute to duarteocarmo/dotfiles development by creating an account on GitHub.
000
Duarte @duarteocarmo.com · 17/09/2026
Invoice dataset is so useful that all the invoices look the same 😂
000
Duarte @duarteocarmo.com · 15/09/2026
I saw a post on LinkedIn saying we should use AMALIA to grade Portuguese high-school exams (the gov. was taking too long) Horrible idea.. But! I benchmarked a bunch of smaller models vs. AMALIA on a bunch of EU-PT benchmarks (including PT-exams) link below
100
Duarte @duarteocarmo.com · 14/09/2026
Learning my new split keyboard is a pain. Send help and your best tips please. This is hell ahahah
000
Duarte @duarteocarmo.com · 10/09/2026
I prefer deepseek v4 flash over pretty much any glm model. @Sentdex you have seen something I have not 😂
000
Duarte @duarteocarmo.com · 08/09/2026
The best thing about LLMs gotten this good is that I can focus on the writing and it will handle all the other bs for me Converting images, generating thumbnails, adding links, animations, visuals ALL THAT BS!!!! (yes it gets the link wrong sometimes, but I catch the mf)
000
Duarte @duarteocarmo.com · 08/09/2026
Cargo bikes, android e-reader tablets, ai-generated text classification and generating great landing pages with LLMs! This month's Retrospectiva is out! duarteocarmo.com/blog/retros...
duarteocarmo.com
Retrospectiva #11
A monthly newsletter about life.
010
Duarte @duarteocarmo.com · 08/09/2026
If I can run 3 deepseek flash sessions in the same time astra is thinking about a single prompt - does this mean I cancel codex? jk
000
Duarte @duarteocarmo.com · 07/09/2026
Building a slide deck today. AGI is here, but I still prefer powerpoint. -> Move logo here -> gray it out -> wtf is this font -> add this image I think it's faster if I just do it myself 😂
000
Duarte @duarteocarmo.com · 06/09/2026
I have to compare two models on a benchmark. Which temp? Which max_tokens? Which top_p setting? Everywhere I look I see "we use the lab recommendation". Ok, so this is the wild west I see.
000
Duarte @duarteocarmo.com · 06/09/2026
I think Casey was right. You can def. replace sleep hours with run hours.
000
Duarte @duarteocarmo.com · 05/09/2026
Evals are hard I admit.
010
Duarte @duarteocarmo.com · 01/09/2026
2024: Use parcel tracking apps 2026: "Check on <tracking_code> and let me know when something changes" Hermes I ❤️ you.
000
Duarte @duarteocarmo.com · 30/08/2026
"Ping!" Agent ready. Running 2-3 agents in parallel doesn't need to be mentally taxing. Pi-jumper allows you to jump between @pidotdev sessions running on tmux with a couple clicks. It's part of pi-tools, some extensions I've built around Pi! Blog post below
100
Duarte @duarteocarmo.com · 25/08/2026
Presenting Bagaço v3 the next generation of Bagaço, the largest pre-training dataset for European Portuguese! ~30 Billion tokens from across the Web, it now also includes Wikipedia and PDF data thanks to the great work from the folks at Hugging Face! Full write up in the comments
110
Duarte @duarteocarmo.com · 19/08/2026
Build a herdr competitor for @pidotdev in 30 seconds today. @ycombinator please invest
100
Duarte @duarteocarmo.com · 10/08/2026
Hiking in the dolomites with an 8 month old, replacing Tmux with herdr, the Laguna technical report, and Tierra Whack. Last month's Retrospectiva is out. duarteocarmo.com/blog/retros...
duarteocarmo.com
Retrospectiva #10
000
Duarte @duarteocarmo.com · 07/08/2026
.@spotify - explain to me why russ is always on my discover playlists? Even though I've clicked hide this multiple times?
000
Duarte @duarteocarmo.com · 07/08/2026
This year I've built evals, classifiers, and even Bagaço, a pretraining dataset for European Portuguese. This raises an obvious question: Can we pretrain LLMs fully in European Portuguese? How do we measure if they're any good? Meet project Ginjinha 🇵🇹 (link in comments)
230
Duarte @duarteocarmo.com · 06/07/2026
Sick home with the flu. GLM-5.2, Scaling Laws, World Cup, and Western States duarteocarmo.com/blog/retros...
duarteocarmo.com
Retrospectiva #9
A monthly newsletter about life.
120
Duarte @duarteocarmo.com · 13/06/2026
It's crazy to me how I just use Matplotlib now to look for flights. But that's just the tip of the iceberg. Here's how I use Saramago, my Hermes Agent duarteocarmo.com/blog/how-i-...
duarteocarmo.com
How I Hermes
Coding on the go, reminders, fitness coaching and more. Here's how I use Saramago, my Hermes Agent.
031
Duarte @duarteocarmo.com · 04/06/2026
A hectic month. Copenhagen Marathon, TranslateGemma, Hoka Hopara, Deep Learning for Biology, and Sakamoto. duarteocarmo.com/blog/retros...
duarteocarmo.com
Retrospectiva #8
A monthly newsletter about life.
001
Duarte @duarteocarmo.com · 26/05/2026
6 months ago I wrote a blog post saying TTS still sucked. The tables are starting to turn. Here's a quick post about options - actual options. duarteocarmo.com/blog/tts-do...
duarteocarmo.com
TTS doesn't suck anymore
6 months ago I wrote a blog post on how TTS still sucked. It's been 6 months, and the tables have turned. We have options now.
010
Duarte @duarteocarmo.com · 12/05/2026
Announcing the second version of Bagaço. 37GB of European Portuguese text. 33 million documents, ~9.3 BILLION tokens. Completely open 😉 duarteocarmo.com/blog/the-la...
duarteocarmo.com
The largest open pretraining dataset for European Portuguese
Announcing Bagaço v2: the largest open pretraining dataset for European Portuguese. 9.7 billion tokens, and a total of 37GB of text, available on Hugging Face.
020
Duarte @duarteocarmo.com · 04/05/2026
This month in Retrospectiva: Using Hermes as a running coach and travel agent, Carlo Rovelli, Jensen Huang, and the Plastic Detox. duarteocarmo.com/blog/retros...
010
Duarte @duarteocarmo.com · 24/04/2026
In Nov 2024, the Portuguese gov. announced a 5.5M EUR investment in project AMÁLIA: a fully open source LLM for European Portuguese. A month ago, the technical report was released. Here's a look at it, and some thoughts on the future of EU-PT LLMs. duarteocarmo.com/blog/amalia...
duarteocarmo.com
AMÁLIA and the future of European Portuguese LLMs
Thoughts on the new technical report from AMÁLIA: The Open Source LLM for European Portuguese
000
Duarte @duarteocarmo.com · 03/04/2026
More Pi. Running and photography apps. Inference engineering and clusters on fire. duarteocarmo.com/blog/retros...
duarteocarmo.com
Retrospectiva #6
A monthly newsletter about life.
000
Duarte @duarteocarmo.com · 25/03/2026
Preparing the Setúbal Half-Marathon and I thought: What if I had a running plan that was actually, you know, for ME. And the calendar export ofc duarteocarmo.com/blog/an-opi...
duarteocarmo.com
An opinionated running dashboard
Some updates to my small and opinionated running dashboard: Tanda Runner
000
Duarte @duarteocarmo.com · 10/03/2026
If you're going to build a European Portuguese LLM, you need a scalable way of detecting it in the wild. Here's a write up of how I got FastText to match SOTA performance with a model that is 10x smaller and 100x faster: It's a bitter lesson. duarteocarmo.com/blog/portug...
duarteocarmo.com
Portuguese variety identification: The bitter lesson
Training a text classifier to detect European Portuguese text at scale. Matching fine-tuned BERT by scaling training data.
000
Duarte @duarteocarmo.com · 01/03/2026
The latest Retrospectiva is out: I talk about Pi, baby carriers, Alphafold, and barbarous distillation attacks duarteocarmo.com/blog/retros...
duarteocarmo.com
Retrospectiva #5
A monthly newsletter about life.
021
Duarte @duarteocarmo.com · 27/02/2026
Today is the day I canceled copilot. github.com/duarteocarm...
github.com
GitHub - duarteocarmo/llama.lua
Contribute to duarteocarmo/llama.lua development by creating an account on GitHub.
010
Duarte @duarteocarmo.com · 23/02/2026
Bagaço (.i.e., Portuguese moonshine) - is the first attempt at a pretraining dataset for European Portuguese. Much inspired by the FineWeb work. Much to do! duarteocarmo.com/blog/bagaco...
duarteocarmo.com
Bagaço: A pretraining dataset for European Portuguese
Bagaço is a first attempt at a pretraining dataset for training LLMs in European Portuguese
020
Duarte @duarteocarmo.com · 17/02/2026
"I don't know how you manage it all" Tmux. Tmux is the answer.
020
Duarte @duarteocarmo.com · 28/01/2026
Italy, Conero national park, running in the hills, agents, and The Pitt. duarteocarmo.com/blog/retros...
duarteocarmo.com
Retrospectiva #4
A monthly newsletter about life.
010
Duarte @duarteocarmo.com · 23/01/2026
You use agents. I use agents. We all use agents. Here's how I police them. duarteocarmo.com/blog/how-to...
duarteocarmo.com
How to police your agents
We're building more than ever. But how do we manage the complexity? We police our agents.
021
Duarte @duarteocarmo.com · 06/01/2026
opencode agent create Go ahead, give it a shot.
000
Duarte @duarteocarmo.com · 05/01/2026
I don't like ads. You don't like ads. We don't like ads. Limpa is like AdBlock + SponsorBlock but for podcasts. GitHub repo below open.substack.com/pub/duarteo...
120
Duarte @duarteocarmo.com · 04/01/2026
Running a web server + a worker used to be a pain. Thanks Claude, didn't know about this one. github.com/nickstennin...
github.com
GitHub - nickstenning/honcho: Honcho: a python clone of Foreman. For managing Procfile-based applications.
Honcho: a python clone of Foreman. For managing Procfile-based applications. - nickstenning/honcho
010
Duarte @duarteocarmo.com · 02/01/2026
Now with agents you can have super meaningful commit messages. It's incredible. Lul
000
Duarte @duarteocarmo.com · 01/01/2026
Django background tasks feedback. Me like.
000
Duarte @duarteocarmo.com · 01/01/2026
complexipy kreuzberg throttled-py batchata I'm living under a rock? Never heard of any of these. Tasty though. tryolabs.com/blog/top-py...
tryolabs.com
Top Python libraries of 2025
Explore our 11th annual Top Python Libraries roundup, featuring two curated Top 10 lists for General Use and AI / ML / Data tools that matter today.
010
Duarte @duarteocarmo.com · 01/01/2026
Dirty little secret. Cancel Claude code, cancel ChatGPT plus. Github copilot is 10 bucks a month. And OpenCode can use them all 🤫
000
Duarte @duarteocarmo.com · 31/12/2025
2000 km for 2025 ✅
000